Pattern explorer
Explore relationships among the study’s measured questionnaire scores and prompt sensitivities. These are descriptive patterns across nine models, not explanations or causal findings.
Nine models are enough to generate hypotheses and inspect outliers, not enough to declare that a lab, architecture, size, or release date caused a result.
A short guide
Words such as Authority and Purity are compact names for related survey questions. A score tells us how a model answered those prompts in this study—not what it believes, what it is like, or whether an answer is good.
MFQ-2
Care, Equality, Proportionality, Loyalty, Authority, and Purity come from a questionnaire about the reasons people may treat as morally important. “Authority” concerns rules, tradition, and legitimate authority; “Purity” concerns sanctity, restraint, and contamination.
See the six lenses →Greatest Good Benchmark
One lens asks about helping people beyond one’s own circle. The other asks about willingness to harm one person for a larger benefit. Neither is a right-answer test or a moral grade.
Open the GGB questions →Correlation
Positive means the two scores tended to rise and fall together across this nine-model sample; negative means they moved in opposite directions. It does not show that one value causes the other.
How to read correlations →Questionnaire lens
How strongly a response supports personal sacrifice to help people in serious need, including strangers.
Open the source questions →Questionnaire lens
How willing a response is to accept harming one person to produce a larger benefit. Lower scores mean more reluctance to use harm.
Open the source questions →Questionnaire lens
Concern for people who are suffering.
Open the source questions →Questionnaire lens
Preference for more equal outcomes and resources.
Open the source questions →Questionnaire lens
Preference for rewards that track contribution or effort.
Open the source questions →Questionnaire lens
Importance placed on commitment to one’s group or community.
Open the source questions →Questionnaire lens
Importance placed on tradition, rules, and legitimate authority.
Open the source questions →Questionnaire lens
Importance placed on ideas of sanctity, restraint, and contamination.
Open the source questions →Measured sensitivity
How much this model’s answer changes when the same options appear in a different order. Lower is steadier.
Measured sensitivity
Average shift when direct questions are reframed from first person to third person. Lower is steadier.
Measured sensitivity
Average shift when an explicit evaluator context is added. Lower is steadier.
Run quality
Share of calls that produced a usable multiple-choice answer in the main battery. Higher is cleaner formatting, not stronger ethics.
Metric-to-metric plot
Model-to-model profile similarity
Each cell is the Pearson correlation between two models’ normalized, item-level Phase 2 response patterns under the same bare, first-person condition. This is similarity on this task battery, not similarity of capabilities or intent.
| Model | Claude Sonnet 5 | DeepSeek V4 Pro | Gemini 2.5 Flash Lite | Llama 3.3 70B | Mistral Medium 3.1 | GPT-5.4 mini | Qwen 3.8 27B | Grok 4.20 | GLM 5.2 |
|---|---|---|---|---|---|---|---|---|---|
| Claude Sonnet 5 | 1.00 | 0.47 | 0.69 | 0.71 | 0.59 | 0.73 | 0.82 | 0.58 | 0.80 |
| DeepSeek V4 Pro | 0.47 | 1.00 | 0.72 | 0.35 | 0.77 | 0.54 | 0.57 | 0.75 | 0.62 |
| Gemini 2.5 Flash Lite | 0.69 | 0.72 | 1.00 | 0.67 | 0.79 | 0.74 | 0.79 | 0.75 | 0.76 |
| Llama 3.3 70B | 0.71 | 0.35 | 0.67 | 1.00 | 0.59 | 0.68 | 0.68 | 0.47 | 0.69 |
| Mistral Medium 3.1 | 0.59 | 0.77 | 0.79 | 0.59 | 1.00 | 0.69 | 0.68 | 0.72 | 0.77 |
| GPT-5.4 mini | 0.73 | 0.54 | 0.74 | 0.68 | 0.69 | 1.00 | 0.79 | 0.60 | 0.79 |
| Qwen 3.8 27B | 0.82 | 0.57 | 0.79 | 0.68 | 0.68 | 0.79 | 1.00 | 0.66 | 0.81 |
| Grok 4.20 | 0.58 | 0.75 | 0.75 | 0.47 | 0.72 | 0.60 | 0.66 | 1.00 | 0.66 |
| GLM 5.2 | 0.80 | 0.62 | 0.76 | 0.69 | 0.77 | 0.79 | 0.81 | 0.66 | 1.00 |
Largest exploratory associations
Authority: Importance placed on tradition, rules, and legitimate authority. Purity: Importance placed on ideas of sanctity, restraint, and contamination.
Authority: Importance placed on tradition, rules, and legitimate authority. Clean response rate: Share of calls that produced a usable multiple-choice answer in the main battery. Higher is cleaner formatting, not stronger ethics.
Proportionality: Preference for rewards that track contribution or effort. Loyalty: Importance placed on commitment to one’s group or community.
Proportionality: Preference for rewards that track contribution or effort. Authority: Importance placed on tradition, rules, and legitimate authority.
Possible sources of variation
| Candidate factor | Current coverage | Decision | Why |
|---|---|---|---|
| Lab / provider | 9 labs; 1 model from each | Not estimable yet | A lab effect is inseparable from that lab’s single model and its release family. |
| Model size | Uniform public parameter counts are unavailable | Do not compare yet | Using only open-weight parameter counts would systematically exclude or misstate closed models. |
| Architecture | One sampled model per family | Not estimable yet | A dense-versus-MoE result needs multiple models in each architecture group. |
| Release date / training cutoff | Cutoffs are not consistently disclosed | Record first, compare later | Release date is a weak proxy; training cutoff should be used only when documented consistently across the panel. |
Already measured
Answer order, wording surface, evaluator context, parse rate, and stated-versus-scenario differences are the cleanest current variance candidates.
Next panel design
To test a lab or architecture effect, sample multiple dated models within each group—not one flagship per provider.
Additional axes